Development and Validation of a Deep Learning Algorithm to Automate Detection of Referable Glaucoma in a Safety Net Teleretinal Program

Abstract

Purpose : Patients of the Los Angeles County (LAC) Department of Health Services (DHS) teleretinal program currently undergo screening for referable glaucoma based on manual grading of fundus photographs by trained LAC DHS optometrists. In this study, we develop and test a deep learning (DL) algorithm to automate the detection process and assess its performance compared to a panel of 12 ophthalmologists.

Methods : Fundus photographs and labels of referable glaucoma were obtained from the LAC DHS teleretinal screening program. Patient-level labels of referable glaucoma were generalized to images from both eyes. A convolutional neural network (CNN) was developed with VGG-19 architecture using separate training and validation datasets. Area under curve (AUC), accuracy, sensitivity, and specificity were calculated to assess algorithm performance using an independent test set graded by 12 ophthalmologists (1 to 15 years of clinical experience). Reference labels were provided by either LAC DHS optometrists (LAC DHS reference labels) or the majority of three glaucoma specialists (expert reference labels).

Results : 12,098 images from 5,616 patients (3,540 referable glaucoma, 2,076 non-glaucoma) were used to develop the DL algorithm. In this dataset, mean age was 56.9 years with 54.9% females and 67.6% Latinos, 9.0% Blacks, 2.7% Caucasians, 6.5% Asians, and 11.2% Other or Unspecified. 1,000 images from 500 patients (250 referable glaucoma, 250 non-glaucoma) were used to test the DL algorithm. In this dataset, mean age was 57.3 years with 52.4% females and 69.2% Latinos, 8.6% Blacks, 5.2% Asians, and 11.8% Other or Unspecified. Algorithm performance (AUC=0.91) exceeded the sensitivity (0.32-0.91) and specificity (0.61-0.98) by all ophthalmologist graders in detecting patient-level referable glaucoma based on LAC DHS reference labels. Performance remained stable (AUC=0.93) and matched or exceeded the sensitivity (0.33-0.99) and specificity (0.68-0.98) of all human graders except members of the expert panel based on expert reference labels.

Conclusions : A DL algorithm can approximate or exceed performance by ophthalmologist graders in detecting referable glaucoma from fundus photographs of LAC DHS teleretinal patients. Implementation of this algorithm in screening workflows could help reallocate eyecare resources and provide more reproducible and timely access to glaucoma care.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.